An AI builder is the profile that builds working product on top of language models: it orchestrates calls to an LLM, designs agents, connects them to the systems that already exist and leaves them evaluated in production. It does not train models from scratch. It assembles them and puts them to work.
That is the definition. The problem is that the word also labels two things that are not a profile.
The most widespread sense does not describe a person: it is a feature of Microsoft Power Platform. The ambiguity is not a vocabulary detail. It changes who you are describing when you open a position, and it changes what the reader understands.
I have sat on both sides of that table: the team asking for an AI builder thinking of whoever assembles the agents, and the candidate convinced they are about to be interviewed on Power Automate. The conversation sorts itself out in five minutes. The badly written job posting costs three weeks.
AI builder labels four different things: a Microsoft Power Platform feature, the category of AI website builders such as Wix or Canva, a professional profile that assembles product on top of already trained models, and the Shakers AI Builder certification, which certifies skills, not a job title. This piece is about the third. That profile orchestrates LLMs, designs agents, integrates them with existing systems and evaluates them before production, and its stack already has measurable demand in the UK: according to the Shakers market analysis, LangChain appears as a requirement in 817 tech postings and the OpenAI API in 862, out of 40,400 active postings between 13 June and 11 September 2026.
Why does AI builder mean four different things?
Because three markets claimed the same word in under three years: a Microsoft product, a category of no-code tools and a job. None of them gave it up. The fourth sense arrived later and does not fight for the same ground: it certifies skills instead of naming a job.
| Sense | What it labels | How to tell it apart |
|---|---|---|
| AI Builder, the product | Microsoft Power Platform feature that adds models to Power Apps and Power Automate | Licensed per credit and lives inside Power Apps and Power Automate |
| AI website builder | Site and automation builders that generate the output with AI: Wix, Canva, Zapier | The subject is a tool and the deliverable is a website or a flow, not a technical decision |
| AI builder, the profile | A person who assembles agents and integrations on top of already trained models | It comes with an article and team verbs: to add, to bring in, to build a team around |
| AI Builder, the certification | Shakers' own standard for senior talent that multiplies its value by orchestrating agents | It does not name a job: it is earned, not held |
The first two describe software. The third describes someone who joins the team and answers for what they deploy. The fourth, the Shakers AI Builder certification, is the standard those skills are certified against before anyone works with a team.
What does an AI builder do inside a technical team?
Four verbs cover the job: orchestrate, integrate, evaluate and deploy.
Orchestrating means chaining model calls with retrieval augmented generation (RAG) and tool use until the whole solves a complete task, which almost never fits in a single call. Integrating is the underrated part: the agent that works in a notebook and the agent that writes to the company CRM are separated by authentication, permissions, rate limits and a written policy for what happens when the model gets it wrong.
Evaluating means keeping a battery of cases that fails when someone touches a prompt and breaks something that used to work. Without that there is no product. There is a demo.
Deploying is the usual work, plus two additions: cost control per call and a way back.
The job exists because of a gap that has been measured. According to BCG's Build for the Future 2025 study (September 2025, n=1,250 companies), only 5% of companies get value from AI at scale and 60% achieve no material value despite having invested. The same report puts at 17% the share of AI value that already comes from agents, with an expectation of 29% by 2028. That gap between what is bought and what reaches production is the job description.
What it does not do also defines the role. It does not train foundation models, it does not own the data platform and it does not replace whoever runs the infrastructure. It works with what already exists.
What stack does it use, and how much demand is there in the UK?
We measured it with our own tech labour market analysis system. Across 40,400 active postings in the UK between 13 June and 11 September 2026, the assembly stack of this profile already sits in the same band as the classic machine learning stack.
| Tool | Postings asking for it | Contrast |
|---|---|---|
| PyTorch | 1,468 | The training benchmark |
| TensorFlow | 1,002 | The training benchmark |
| OpenAI API | 862 | The most requested LLM API |
| LangChain | 817 | Close behind TensorFlow |
| Anthropic API | 561 | Ahead of Ruby (409) |
| Hugging Face | 215 | Open models |
The reading is in the right hand column. In the UK the training stack still leads, but the gap has closed within a single hiring cycle: the OpenAI API and LangChain already share the band that TensorFlow owned alone, and Anthropic, an API that barely existed three years ago, appears in more postings than Ruby. The direction of travel is clear: more postings ask teams to assemble models than to build the plumbing around training pipelines.
Two warnings about the method. The counts overlap and cannot be added up, because a posting that asks for LangChain and the OpenAI API counts in both rows. And the term ai engineer appears in 2,343 postings of the period, which is not the same as saying there are 2,343 ai engineer roles: it shows up in the text, sometimes only in passing.
How is it different from an AI engineer or a prompt engineer?
In the point of the cycle where they come in, and in what they deliver.
| Profile | Focus | Deliverable | Maturity of the term |
|---|---|---|---|
| AI builder | Assembling product on top of already trained models | Agent or feature in production | Emerging, no agreed description |
| AI engineer | Full system cycle, including serving and scaling the model | Sustainable inference platform | Established, 2,343 mentions in the measured period |
| Prompt engineer | Design and evaluation of instructions | Prompt library with metrics | A skill more than a job: it appears inside the other two |
| Classic developer | Product with no model component | Application | Consolidated |
The frontier with the AI engineer is porous and depends on size. In a fifty person team it is the same person. Past a certain scale they split: sustaining the inference platform and building on top of it stop fitting in the same week.
With the classic developer the frontier is sharper. Whoever calls an OpenAI API is not yet an AI builder; they become one when they have to answer for why the model replied that, and what changed so it does not happen again.
What should a team look at before working with an AI builder?
Start by deciding which of the four verbs it is missing. Most teams that open this position need integration and evaluation, and then write a posting that talks about orchestration.
Telling apart who has genuinely done that is where almost every process falls down. A list of tools does not discriminate: this stack is learned in tutorials with an ease the training stack never had.
The question that does discriminate is uncomfortable and cannot be rehearsed: which of your agents reached production, what did it get wrong in its first week, and what did you change so it stopped. Whoever has been there answers with a concrete case in seconds. Whoever has not answers by talking about the model.
That is the criterion we apply when validating profiles. Talent certified in AI skills means evidence of execution, not a declared repertoire, and it is the difference between a list of candidates and a hiring infrastructure you can lean on.
AI Builders sets out what is certified before this profile works with a team. If your question sits one step earlier and you are still weighing whether you need an agent, or which kind, the groundwork on AI agents and on agents for companies covers both.
Frequently asked questions about the AI builder
- Is an AI builder the same as Microsoft's AI Builder?
No. Microsoft's AI Builder is a Power Platform feature that adds pre-trained models to Power Apps and Power Automate: it is licensed per credit and never leaves the Microsoft environment. An AI builder is a person who assembles agents and integrations on top of third party models and answers for what they deploy. What separates them is that one is bought and the other joins the team, not how they are written.
- Is an AI builder an AI website builder?
Not that either. AI website builders, such as Wix or Canva, generate a site or an automation from an instruction: the subject of the sentence is the software. In the profile sense the subject is a person, and their deliverable is not a website, it is an agent integrated with the company's systems.
- What is the difference between an AI builder and an AI engineer?
The first comes in on top of models that already exist and delivers the agent or the feature in production. The AI engineer covers the full system cycle, including serving and scaling the model, and delivers the inference platform. In small teams it is the same person; past a certain scale they split.
- What skills does an AI builder need?
LLM orchestration with LangChain or equivalents, use of the OpenAI and Anthropic APIs, RAG for context retrieval, agent design with tool use and automated evaluation. On top of that, what almost never makes it into the posting: authentication, permissions, cost control per call and a way back when the model fails.
- Is there demand for AI builders in the UK?
Of the stack, yes, and it is measurable. According to the Shakers market analysis, across 40,400 active tech postings in the UK between 13 June and 11 September 2026, the OpenAI API appears in 862 and LangChain in 817, both inside the band TensorFlow (1,002) owned alone until recently, and Anthropic's in 561, ahead of Ruby (409). The counts overlap and must not be added up.
- How does a company work with an AI builder?
First it decides which of the job's four verbs it is missing, because orchestrating, integrating, evaluating and deploying attract different candidates. Then it asks for evidence of execution rather than lists of tools: which agent reached production, what it got wrong and what was changed.
- What is the Shakers AI Builder certification?
The AI Builder certification is Shakers' own standard for senior talent that multiplies its value by orchestrating agents. It does not describe a job: it certifies skills, AI maturity and demonstrated capability, with profile review, practical tests and signals of real agent use. It gives companies more reliable signals before anyone works with a team.
